Google Unveils DiffusionGemma to Explore Text Diffusion

Episode 1 · DiffusionGemma: Parallel Diffusion Decoding Breaks LLM Inference Speed Bottleneck (2026-08-04, 8 posts)

Google DeepMind released the technical report for DiffusionGemma, an experimental open-weight language model. Using discrete diffusion, it breaks the traditional autoregressive token-by-token decoding paradigm, enabling parallel iterative refinement of 256-token blocks and achieving 1500 tokens per second on a single GPU. This offers a novel parallel decoding approach to overcome LLM inference speed bottlenecks.

Confirmed

• Core mechanism and parameters: DiffusionGemma is built on Gemma 4 (25.2B total parameters, 3.8B active), using discrete diffusion with bidirectional attention and iterative refinement to output 256 tokens in parallel.
• Online post-training: After supervised fine-tuning (SFT), the team combined sampler distillation with reinforcement learning. This online post-training mechanism pushes the model's quality and inference speed to a new Pareto frontier.
• Performance advantages: According to author @bodonoghue85, the architecture not only achieves 1500 tokens/sec on a single GPU but also shows significant latency and throughput advantages at low-to-moderate batch sizes, with benefits increasing with context length.
• Application scenarios: In strongly input-constrained scenarios (e.g., code editing where output is highly similar to input), the model shows significant advantages over autoregressive models.

Why it matters

• This 55-page technical report thoroughly explores the application of diffusion models in language generation. Its parallel decoding idea shows great potential in tasks requiring highly constrained or structured outputs, potentially greatly improving inference efficiency of large models in real-world engineering.

• DeepMind Releases DiffusionGemma: Discrete Diffusion for Ultra-Fast Text Generation — deepmind · 2026-08-04
• Google's DiffusionGemma: Parallel Decoding Breaks LLM Text Generation Speed Limits — bodonoghue85 · 2026-08-04
• DiffusionGemma Tech Report: Parallel Decoding Breaks LLM Inference Speed Limits — bodonoghue85 · 2026-08-04
• Text Diffusion Models Offer Latency and Throughput Gains at Medium Batch Sizes — bodonoghue85 · 2026-08-04
• DiffusionGemma Outputs 256 Tokens Simultaneously, Accelerating Code Editing — bodonoghue85 · 2026-08-04
• DiffusionGemma Tech Report: Online Post-Training Pushes New Quality-Speed Pareto Frontier — bodonoghue85 · 2026-08-05
• DiffusionGemma Report: Parallel 256-Token Generation Breaks AR Bottleneck — SungjinAhn_ · 2026-08-05
• DiffusionGemma Technical Report: Parallel Text Gen at 1,500 Tokens/sec on Single H100 — kastnerkyle · 2026-08-06

Episode 2 · Advantages of Text Diffusion Models: Non-Causal Correction and High Throughput (2026-08-04, 2 posts)

Text diffusion models outperform traditional autoregressive models by leveraging bidirectional attention and iterative refinement for non-causal error correction. Additionally, they offer significant latency and throughput advantages in small to medium batch sizes, which scale positively with increased context length.

• Text Diffusion Models Show Latency and Throughput Gains for Longer Contexts — bodonoghue85 · 2026-08-04
• Diffusion Models' Edge: Bidirectional Attention and Iterative Refinement Enable Non-Causal Error Correction — bodonoghue85 · 2026-08-04

Episode 3 · Google Releases DiffusionGemma Technical Report on Text Diffusion (2026-08-11, 2 posts)

Google has published the full technical report for DiffusionGemma, sharing key insights into text diffusion models. Concurrently, the llama.cpp community has already submitted a PR to adapt the model for local inference.

• Google Releases DiffusionGemma Technical Report on Text Diffusion — rseroter · 2026-08-11
• DiffusionGemma Technical Report Released, llama.cpp Adaptation Underway — pmttyji · 2026-08-11